Privacy set intersection (PSI)

What is Privacy set intersection (PSI)?

Privacy set intersection (PSI) is a cryptographic protocol that allows two parties to identify the common elements between their private datasets without either party revealing any information about the elements that are not in the intersection. Each organization learns only what records overlap, never the full contents of the other party's set. PSI enables privacy-safe data collaboration across advertising, measurement, and identity resolution use cases.

How it works

PSI works by applying cryptographic transformations to each party's dataset so that comparisons can be performed on encoded values rather than raw data. The protocol proceeds through three core phases.

Preprocessing

Each party independently applies a cryptographic function, such as a hash or an encryption scheme, to every element in their private set. These transformed representations conceal the actual values while preserving the mathematical properties needed for comparison. Common techniques include oblivious pseudorandom functions (OPRFs), Diffie-Hellman key exchange variants, and garbled circuits.

Comparison

The two parties exchange their cryptographic representations and perform a secure comparison. The comparison is structured so that only matching elements produce a recognizable signal. Neither party can reverse-engineer the other party's non-matching values from the exchanged data. Secure multi-party computation and homomorphic encryption are frequently used to enforce this property.

Output

The protocol produces either the size of the intersection, a list of matched identifiers in encrypted form, or an aggregated signal derived from the overlap. The output format depends on the use case. For audience measurement, the result might be a count of overlapping users. For identity resolution, it might be a set of pseudonymous matched identifiers that neither party can link back to raw personal data.

The security guarantees of PSI depend on the strength of the underlying cryptographic primitives and the correctness of the implementation. A properly implemented PSI protocol provides provable privacy guarantees under well-defined threat models.

Why it matters

PSI addresses a fundamental tension in digital advertising and measurement: organizations need to collaborate on data to deliver accurate attribution and audience insights, but sharing raw user-level data creates significant privacy and regulatory risk. Regulations such as GDPR and CCPA restrict how personal data can be transferred between parties, making traditional data-sharing approaches legally and operationally complex.

With PSI, two parties, such as an advertiser and a publisher, can determine how many users they have in common without either side gaining access to the other's customer records. This enables use cases such as reach and frequency measurement, audience deduplication, cross-platform attribution, and clean room analytics, all without exposing individual-level data.

PSI is increasingly relevant as third-party identifiers such as cookies and mobile advertising IDs become less available due to platform restrictions and user opt-outs. It provides a durable alternative that operates on first-party data and satisfies privacy requirements by design. For mobile measurement specifically, PSI can support conversion matching workflows that comply with user consent frameworks while still producing actionable campaign signals.

How to implement privacy set intersection for mobile measurement

Implementing PSI for mobile measurement requires coordination between technical, legal, and data teams. The following steps outline a practical approach.

Define the use case and data sets. Determine which parties will participate in the PSI computation and what data each party holds. Common scenarios include matching advertiser CRM records against publisher audience data, or reconciling attribution signals between an MMP and an ad network.

Select a PSI protocol variant. Choose a protocol that fits your performance and security requirements. OPRF-based PSI is widely used for its efficiency at scale. Homomorphic encryption-based approaches offer stronger privacy guarantees but require more computational resources. Many privacy-preserving technology vendors and data clean room providers offer pre-built PSI infrastructure.

Establish a data governance agreement. Both parties must agree on what the output of the PSI computation will be, how it will be used, and how long any derived signals will be retained. This agreement should be documented and aligned with applicable data protection regulations.

Standardize identifiers before encoding. Normalize the identifiers in each party's dataset, such as email addresses or phone numbers, before applying cryptographic transformations. Inconsistent formatting will reduce match rates and produce inaccurate intersection results.

Validate the output before acting on it. Verify that the intersection size and any derived signals are statistically plausible. Anomalously large or small intersections may indicate data quality issues or implementation errors rather than true audience overlap.

Integrate with downstream measurement workflows. Feed PSI outputs into attribution pipelines, campaign reporting, or audience suppression lists as appropriate. Platforms like Airbridge support privacy-preserving measurement integrations that can incorporate PSI-derived signals alongside other attribution data.

Related concepts